Deep learning-based pathology signature could reveal lymph node status and act as a novel prognostic marker across multiple cancer types

Abstract

Background: Identifying lymph node metastasis (LNM) relies mainly on indirect radiology. Current studies omitted the quantified associations with traits beyond cancer types, failing to provide generalisation performance across various tumour types.

Methods: 4400 whole slide images across 11 cancer types were collected for training, cross-verification, and external validation of the pan-cancer lymph node metastasis (PC-LNM) model. We proposed an attention-based weakly supervised neural network based on self-supervised cancer-invariant features for the prediction task.

Results: PC-LNM achieved a test area under the curve (AUC) of 0.732 (95% confidence interval: 0.717-0.746, P < 0.0001) in fivefold cross-validation of multiple cancer types, which also demonstrated good generalisation in the external validation cohort with AUC of 0.699 (95% confidence interval: 0.658-0.737, P < 0.0001). The interpretability results derived from PC-LNM revealed that the regions with the highest attention scores identified by the model generally correspond to tumours with poorly differentiated morphologies. PC-LNM achieved superior performance over previously reported methods and could also act as an independent prognostic factor for patients across multiple tumour types.

Discussion: We presented an automated pan-cancer model for predicting the LNM status from primary tumour histology, which could act as a novel prognostic marker across multiple cancer types.

Fig. 1. The analysis pipeline and the neural network architecture of this study.

Our computational approach captures informative representations from the 768-dimensional features through attention-based aggregation to extract spatial context from the local morphology. Then, an attention-based aggregation means sum over all instances in a bag to get a global feature for classification, and the fused features are fed into a classifier to calculate the cross-entropy loss.

Fig. 2. The overall performance of the PC-LNM.

a Receiver operating characteristic curve analysis evaluated the performance of the PC-LNM in the cross-validation cohort. b Receiver operating characteristic curve analysis evaluated the performance of the PC-LNM in the external validation cohort. PC-LNM pan-cancer lymph node metastasis, AUC area under curve, CI confidence interval.

Fig. 3. Interpretability of the deep learning-based pathological model.

WSI whole slide image, ROI region of interest, TMB tumour mutational burden, TMB-H TMB-high, TMB-L TMB-low.

Fig. 4. Prognosis prediction through the PC-LNM.

a Kaplan–Meier survival analysis stratified by PC-LNM for overall survival in the TCGA cohort. b Kaplan–Meier survival analysis stratified by PC-LNM for overall survival in the CPTAC cohort. PC-LNM pan-cancer lymph node metastasis, HR hazard ratio, CI confidence interval TCGA The Cancer Genome Atlas, CPTAC Clinical Proteomic Tumour Analysis Consortium.